Research On Spatial And Temporal Distribution Prediction Model Of Urban Invasion Cases Based On Deep Learning

With the development of society and economy,criminal acts such as theft,looting,robbery and fraud,which are collectively referred to as crimes against property by the public security organs,are showing a high incidence,which seriously affecting the sense of security and happiness of people's lives.In fact,the spatial and temporal distribution of crimes against property is not random or uniform,instead they present spatial and temporal clustering,spatial-trmporal interaction,and they happen in near repeat.The prevention and control of crimes against property have always been one of the important tasks of the public security department,therefore,based on the law of spatial and temporal character of crimes against property,predicting the spatial and temporal distribution on a micro time scale is significantly important for public security management and patrol plan,thus effectively combat crime.At present,the crimes against property prediction based on the law of spatial and temporal distribution is widely studied,but the spatial and temporal distribution characteristics of crime on the micro time scale are difficult to extract,which becomes the bottleneck problem of crime prediction,and restricts the crime prediction on a smaller time scale;however,the deep learning algorithm is excellent in image feature recognition and it is also suitable for the extrraction of sparse case distribution features.Therefore,the crime prediction combined with deep learning algorithm has become one of the hot trends in recent years.This paper takes crimes against property as the research object,and predicts the spatial and temporal distribution on hourly basis,by research on impact factors for the spatial and temporal distribution of crimes against property,a spatial and temporal distribution prediction model for crimes against property is constructed based on Deep Spatio-Temporal Residual Networks(ST-ResNet),a crime prediction result evaluation method is estabilished,which can comprehensively analyze and evaluate the prediction results.The research content and achievements of this paper are as follows:(1)Taking the crimes against property as the research object,through statistical analysis of the crimes against property from the perspective of time and space,the time-space clustering and periodicity of the crimes against property in the study area are obtained;then,by correlation analysis of the amount of crimes against property and different environmental impact factors,the environmental impact factors related to the crimes against property were determined;the above provides the theoretical basis for the prediction model.(2)Based on the near repeat phenomenon of crimes against property,this paper studies the fusion method for parameter setting mechanism of ST-ResNet model and crimes against property distribution characteristics at different moments,and establishes a spatio-temporal distribution prediction model based on ST-ResNet,this solves the difficulty in feature extraction for sparse case distribution,and realizes the space-time distribution prediction on hourly basis.(3)Through the fitting of predictin model and analysis of the accuracy for crimes against property prediction,and by the comparison with the real cases in different time and different areas,the process from model fitting to real data analysis,from the overall to local crime prediction results is realized.The research comprehensively evaluated the prediction results of ST-ResNet model,the results show that the ST-ResNet prediction model is better than the ST-CNN prediction model,and the prediction accuracy of the ST-ResNet model varies by temporal and regional differences.This paper takes the crimes against property in the core area of Suzhou as an example.Through comparison and analysis,under the condition of 2.4km spatial resolution and 2 hours time resolution,the city-wide prediction accuracy of ST-ResNet model is 41%,and the prediction accuracy is better than ST-CNN,which is 35%_In addition,it is found that the prediction accuracy based on ST-ResNet model is not the same at different time and different regions.Suzhou Gusu District is of high case density,the average prediction accuracy here in 2013 is as high as 60%,and the highest monthly average prediction accuracy reaches 75%;Suzhou Huqiu Disrict is of the low case density,the average prediction accuracy is only 10%in 2013,and the prediction accuracy rate varies from time to time,the average forecast accuracy in the early morning of 2013 is around 15%,the prediction accuracy in other time period is about 50%.In summary,the overall prediction accuracy of crimes against property in Suzhou city is about 41%,and there are differences in different time and region,but the overall predicting results are good,which can provide decision and support for the public security prevention and control work in Suzhou.